JEV GUIDE

Jev writing quality checks for Markdown and technical docs

“Make this more engaging” gives an author little to act on. slop-grader evaluates explicit rules and reports lines or document dimensions that deserve attention. The author or another writing tool then revises the text. This is not evidence of who wrote a document.

Jev Gallery · Reviewed 2026-09-20

Make writing requirements reviewable

Start with the document’s job. A setup guide needs prerequisites and executable steps; an opinion piece may rely on rhetoric and personal voice. Translate vague preferences into observable questions, such as whether a requested action identifies a target or a numerical claim has support.

Prefer deterministic checks for broken links, required headings and code syntax. Reserve semantic evaluation for requirements that need interpretation.

Separate line checks from document scores

The reviewed grader constructs Noul questions for line rules and Score questions for document rules. A line finding can locate a specific problem; a document score can help assess structure, prerequisites or task coverage.

A report is input to revision, not a revised manuscript. If a sentence lacks a fact, the author must supply it. A generation tool can propose wording but should not invent the missing evidence.

Source: slop-grader — grader.ts ↗

A custom rule design

This editorial Markdown fixture follows the project’s documented format. It has not been executed through the syntax checker or a live model. The rule targets an actionable documentation problem rather than inferring AI authorship.

Markdown · editorial example

# Line Rules

## unexplained_action
Does the line ask the reader to change a setting without identifying what to change?

### Criteria
- **true**: A setting change is requested, but no setting or value is identified.
- **false**: The target is specified, or the line does not request a setting change.

# Document Rules

## prerequisites
How clearly does this setup guide state what the reader needs before starting?

### Criteria
- Required inputs and access are absent
- Some requirements are stated, but important ones are missing
- Required inputs and access are identifiable
- Requirements and a way to verify them are explicit

Source: slop-grader — README.md ↗

Calibrate with positive and negative fixtures

These synthetic cases contain no invented model scores. Have an editor label them first, then investigate whether disagreements come from missing context or unclear criteria.

Calibrate with positive and negative fixtures
Synthetic passage	Review target	Editorial response
“Optimize the relevant settings.”	No setting is identified	Ask for the field and value
“Set timeoutMs to 0 for one evaluation.”	Specific target and value	Do not flag mere brevity
“This increases efficiency by 90%.”	Unsupported measurement	Request a source and conditions
A prohibited phrase inside a quotation	Quoted context differs from narration	Review rather than automatically replace

Close the loop without losing meaning

Validate rule syntax before grading a permitted draft. The project’s --check runs without model inference: it does not establish the rule’s accuracy. Keep the draft, rules, model identifier and report, then review false positives before applying edits.

Recheck the original findings and affected passages after revision. Compare names, numbers, citations and conclusions with the original. Do not remove necessary limitations to improve a score. Real grading sends text to the selected provider; the author should control that input scope.

Source: slop-grader — README.md ↗

Common questions

Can it remove all AI-sounding prose? It checks supplied style criteria, not authorship, and cannot guarantee every reader’s preference.

Does this establish Chinese-language quality? No. Custom rules are possible, but bundled English or German grammar rules do not validate Chinese performance. Build and label language-specific examples.

Should every finding block CI? Begin with advisory reports and review false positives before promoting stable, explainable checks to gates. We reviewed source code and did not ask Jev to grade a manuscript.